1. Laboratory of Environmental Geosimulation (LEDGE), Geography Department, University of Montreal, 520 Chemin de la Côte-Sainte-Catherine, H3C 3J7, Quebec, Canada.*Corresponding author
2. Centre d’étude de la forêt, Département des sciences biologiques, Université du Québec à Montréal, C.P. 8888, Succursale Centre-Ville, Montréal, QC H3C 3P8, Canada.
Abstract
Wildfires are the main cause of forest disturbance in the boreal forest of Canada. Climate change studies forecast important changes in fire cycles, such as increases in fire intensity, severity, and occurrence. The geographical information system (GIS) based cellular automata model, BorealFireSim, serves as a tool to identify future fire patterns in the boreal forest of Quebec, Canada. The model was calibrated using 1950-2010 climate data for the present baseline and forecasts of burning probability up to 2100 were calculated using two RCP scenarios of climate change. Results show that, with every scenario, the mean area burned will likely increase on a provincial scale, while some areas might expect decreases with a low emission scenario. Comparison with other models shows that areas forecasted to have an increase in fire likelihood, overlap with predicted areas of higher vegetation productivity. The results presented in this research aid identifying key areas for fire-dependent species in the near future.
Highlights
Climate Change will likely alter wildfire patterns in the boreal forest of Quebec BorealFireSim can model those wildfire patterns
Forecasts show increases in fire likelihood
Important changes in fire patterns for the northeastern part of the boreal forest Fire likelihood and vegetation productivity forecasts are overlapping
Keywords:
47
Introduction
Fire is the main source of disturbance in the boreal forest of Quebec (Natural Resources Canada, 2014). Wildfires are essential for forest regrowth of tree species such as jack pine or black spruce that depend on extreme heat to reproduce, as well as for insects or bird species depending on dead trees, or snags (Bonnot et al., 2009; Nappi et al., 2003). For the last decade, fire suppression costs in Canada ranged from 500 million to 1 billion dollars per year and burned more than 2.3 million hectares annually (Natural Resources Canada, 2014). Moreover, natural wildfires burning more than 200ha account for 97% of the total area burned across Canada and represent 3% of the wildfires (Natural Resources Canada, 2014). Wildland fires are caused mainly by four factors: fuels, climate- weather, ignition agents, and people (Flannigan et al, 2009). Numerous studies have concluded that upcoming climate change will be a major driver of ecological change (Flannigan et al., 2000; IPCC, 2013). Dale et al. (2008) states how the interactions between climate, disturbances and forest systems are critical to determine climate change impacts on forests. ). Among the biological impacts of climate change, variations in migration patterns of animals, increasing prevalence of wildfires and massive insect outbreaks are the most relevant (IPCC, 2013) Moreover, Mantyka-Pringle et al. (2015) demonstrated, in a study on the interplay between climate change and land-cover change, that adding climate change to land-cover change could increase the impacts of land-cover changes by up to 43% for birds and 24% for mammals. Fire frequency, size and seasonality would likely also be affected by climate change (IPCC, 2013). ). Keane et al (2008), showed that predicted future climate change will likely cause major shifts in landscape vegetation dynamics and this shift is likely to be enhanced by independent changes in biophysical conditions. Changes in fire behavior will affect forest value for wildlife habitat as well as for the industry. Additionally, fire ignition and spread depend on the amount and frequency of precipitation, the type of forest cover and different conditions, such as thunderstorms, topography and wind speed, amongst others; thus these variables should be included within wildfire models (Dale et al., 2001).
Modelling fire behaviour and spatiotemporal patterns enable better understanding of the feedbacks and interactions occurring in forested landscapes. Fire propagation models are usually deterministic and based on linear statistics; examples of such type of models are the Canadian Wildland Fire Effects Model (CanFIRE) (de Groot, 2012). CanFIRE is used by the Canadian Forest Service to predict the physical and ecological impacts of fires. Another widely used model is FARSITE (Finney, M.A., Andrews, 1999), a GIS-based fire growth model that is used to produce maps of fire behavior on a fire
48
event. Even though FARSITE is extremely powerful and couples statistical decision making to GIS, this model is not meant for large scale spatiotemporal fire dynamics, but for fire spread across landscapes (Finney, M.A., Andrews, 1999). While FARSITE and other FARSITE-based models like Fire-BGC (Green et al., 1995) – a spatially-explicit fire succession model designed to investigate long-term trends in landscape pattern under historical and future fire regimes – focus on fire spread across a landscape, BorealFireSim works at a provincial scale, and focuses on a long term spatio-temporal changes in wildfire patterns in the boreal forest, dealing with many fire events in space and in time (Finney, M.A., Andrews, 1999; Keane et al., 1999). Given that wildfire ignition can be caused by diverse interacting conditions, such as climate, elevation, dryness, tree species, weather, and presence of wet areas, the complex dynamics between these conditions give rise to spatial patterns of burned areas, emerging from local interactions to global scale patterns through time. The dynamic behaviour of wildfire processes can be studied by complex systems theory, which takes into account non-linearity of processes and feedbacks with the environment. The term complexity is used in this research, to represent the process by which identical initial conditions in an environment will give rise to different outcomes if the experiment is repeated multiple times (Batty and Torrens, 2001). Researchers often integrate complex behaviours into simple models using stochastic and dynamic modelling approaches. Among these approaches, cellular automata (CA) models have been proven effective to reproduce non- linear processes (Wolfram, 1994).
Cellular automata are models comprising a grid of cells where each one has a finite number of states. The state of a cell is influenced by the neighboring cells via transition rules. These transitions rules are applied to each cell for a certain number of time steps. In CA models, the state of a cell can be summarized with the following equation:
𝑆𝑖,𝑗(𝑡 + 1) = 𝑓{𝑁𝑖,𝑗(𝑡), 𝑆𝑖,𝑗(𝑡), ∆𝑇} (3)
where the state (S) of a cell i,j at a time (t+1) is a function of its neighborhood 𝑁𝑖,𝑗(𝑡), and its state at
the previous time 𝑆𝑖,𝑗(𝑡) within a discrete time step ∆𝑇. The major advantage of this approach is that
instead of running simulations on the whole system with complex mathematical equations, simple rules are imposed on cells that can only interact with their neighbors. During and after the simulation, spatial patterns emerge from these local interactions between cells (Li and Magill, 2001).
The approach used in this research is a GIS-based cellular automata, which allows us to model dynamic, complex and non-linear interactions on large spatial and temporal scales. When coupled to
49
GIS, CA models make powerful tools for simulating complex spatiotemporal phenomena such as wildfire. While most of cellular automata models represent abstract or virtual environments, adding actual georeferenced map layers lets us model complex dynamics taking into account real landscapes and that is especially why GIS-based CA have been used in numerous fields. Examples of CA and GIS- based CA models can be found in multiple studies of dynamic processes such as land usecover change and urban dynamics (Kocabas & Dragicevic, 2006, 2007; Ward et al., 2000; Yeh & Li, 2003), invasive species (Bone et al., 2006; Perez & Dragicevic, 2012) and forest fires (Alexandridis et al., 2008, 2011; Yassemi et al., 2008), to cite only few of the applications. In forest fires studies, Alexandridis (2011) showed the power of GIS-based cellular automata combined with meteorological data as a way to efficiently predict the evolution of fire front on forest landscapes. Alexandridis (2011) also included the spotting effect which is a phenomenon where burning material is transported by wind to areas not adjacent to the fire front, sometimes causing the ignition of a new, independent, fire event. Even though spotting could be important for fire front evolution models, this phenomenon is not relevant on a provincial scale, where the spatial resolution doesn’t allow these short range (100 meters approximately) dynamics.
This research presents a novel GIS-based CA modeling approach named BorealFireSim, where the importance of model variables and transition rules are based on literature and on a thorough sensitivity analysis. Moreover, the BorealFireSim uses province wide information to simulate the probability of annual forest wildfires under current and future climate scenarios. In general, fire models consist of fire front spread models, that is, the evolution of a single fire depending on various variables, such as bush density, bush flammability and wind speed and direction (Li & Magill, 2001). Alternatively, BorealFireSim model aims to simulate probable fire patterns on a provincial scale under climate change and does not take into account fine scale fire spread behaviour as much as a fire front model would.
The main goal of this study is to simulate the complexities of wildfire processes and to identify changes in their spatial patterns throughout Quebec’s boreal forest. Fire spread and post-fire regrowth behaviour in BorealFireSim are based on climate and environmental variables. Furthermore, after calibrating the model to reproduce current fire patterns, simulations are made using future climate data (CMIP5) for different representative concentration pathways (RCPs) (IPCC, 2013; Hijmans et al., 2005). The following manuscript is organized in five sections. The material and methods (input data, model’s transition rules, model flowchart, variables, algorithms and outputs) are detailed in the
50
following section. The results and discussion, section three, show maps of burning chance, cell states after 100 years, fire risk maps and statistical analysis. Likewise, results are compared to forecasted dynamic habitat index maps for 2050 and 2080 (Nelson et al., 2014) to identify relevant areas for species conservation. Finally, limitations and conclusions come back on the assumptions and limitations of BorealFireSim and summarize the model results.